Machine learning classification for tool life modeling using production shop-floor tool wear data

Abstract

Abstract Tool wear is an important limitation to machining productivity. Tool wear in machining is difficult to predict due to large number of influencing variables and tool-to-tool performance variation. As a result, empirical models (such as Taylor’s tool life equation) or physics-based models require experimentation to calibrate model coefficients, which is infeasible in an industrial setting due to large number of tool-material combinations. In this paper, machine learning classification methods for modeling tool life using production shop-floor tool wear data are presented. Tool wear data is simulated using a known ‘true’ tool life curve as a function of cutting speed. Two machine learning classification methods are evaluated, Support Vector Machines, and logistic, and the results are compared to the true tool life curve. Results show good agreement with the true tool life curve using the radial basis function, and polynomial kernels of Support Vector Machine and logistic classification with log transformation of input variables. The classification accuracy increases with the number of training data points. A method to generate synthetic data to augment sparse and unbalanced datasets is presented. Results show that domain knowledge and experience can be effectively used to generate synthetic data to improve tool life classification. The proposed machine learning classification approach offers a powerful and practical approach for modeling tool life in machining in an industrial setting.